Newsletter · 2026-07-18 · Ashutosh Agarwal

Chips Fall Into a Bear Market, SK Hynix Debuts, TSMC and ASML Beat - Semiconductor Podcast Briefing - Week of July 17, 2026

The main basket of semiconductor stocks fell into a bear market, SK Hynix pulled off the largest-ever US listing by a foreign company before sliding back below its debut price, TSMC and ASML both beat and both sagged, and IBM's record 25% drop set off inventory-glut fears. Our synthesis of the semiconductor podcast tape for the week ending July 17, 2026.

Semiconductor Podcast Briefing

Week of July 17, 2026: Chips Fall Into a Bear Market, SK Hynix Debuts, TSMC and ASML Beat

This was the week the great 2026 chip rally cracked. After a run that saw the main basket of semiconductor stocks (the Philadelphia Semiconductor Index, or "SOX") roughly double from late March to late June, the group tipped into what traders call a bear market, a fall of more than 20% from its high. Memory-chip makers fell even harder. The podcasts this week were dominated by one question asked from a dozen angles: was this a healthy pause in a real, multi-year boom, or the first crack in a bubble? Below is what the most-followed investors, analysts, and industry insiders actually said, with the numbers and the names.

A note on jargon, since a lot of it flew around this week: HBM ("high-bandwidth memory") is the ultra-fast, expensive memory that sits right next to an AI chip; DRAM is standard computer memory; NAND is flash storage. Capex is capital spending, the money hyperscalers (giant cloud companies like Google, Amazon, Microsoft, Meta) pour into data centers. WFE is "wafer fab equipment," the machines that make chips. A foundry is a factory that manufactures chips for others (TSMC is the biggest). I define the rest as they come up.

TL;DR, Things That Mattered This Week

• Chips entered a bear market. On The Joe Reis Show (July 17), host Joe Reis summed it up: "The Philadelphia Semiconductor Index fell roughly 24% from its June peak. It's in a bear market right now," after having "ridden approximately 83% during 2026." On Squawk on the Street (July 17), CNBC's Mike Santoli put numbers on it: "Semis as a group down almost 20%, memory down more than 30% from the highs."
• The trigger was a cheap Chinese AI model. A Chinese lab called Moonshot released a new model ("Kimi K3") that, per CNBC's Pippa Stevens (July 17), briefly "taken the number one spot" on a coding leaderboard and was pitched as "as powerful as the Anthropic and OpenAI models", reviving fears that American companies won't need to spend as much on chips if good AI gets cheap.
• SK Hynix had one of the wildest debuts in memory. It raised about $26.5 billion on the Nasdaq on July 10 (the largest-ever US listing by a foreign company, 7x oversubscribed with ~$200B of demand), priced at $149 per share-equivalent, opened up ~14%, then suffered a record one-day plunge back home in Korea on July 13, and by July 17 had fallen back below its $149 debut price.
• TSMC and ASML both beat and both sagged. TSMC (July 16) grew profit 77% year-over-year with gross margins of 67.7% and pledged another $100 billion for US factories (total now $265 billion), and the stock still fell. ASML (July 15) raised its full-year outlook for the second time this year and lifted gross-margin guidance from 52% to 55%, and barely moved.
• The scariest single data point came from IBM. IBM fell 25% in a day, its worst on record, wiping out $68 billion, after saying clients had rushed to buy memory and servers ahead of price hikes, hurting its own results. Several hosts called it the "canary in the coal mine" for a coming inventory glut.
• The bull-bear fight got sharp and specific. Bulls (Gavin Baker, Ben Pouladian, Janus Henderson's Ken Fish) argue this is a real shortage lasting into 2028-2030, protected by new multi-year contracts. Bears (Jim Chanos, Peter Boockvar, Louis-Vincent Gave) argue hyperscaler returns are falling fast, memory margins are a cyclical top, and Chinese competitors are coming for the 85% profit margins.
• China's memory champions got named repeatedly. CXMT (China's #4 DRAM maker) IPO'd in Hong Kong the same week, and multiple guests warned that CXMT and Yangtze Memory (YMTC) are the real long-term threat to Micron, Samsung, and SK Hynix.

1. AI Chip Demand and Hyperscaler Capex (NVDA, AMD, AVGO, MRVL)

The demand debate split cleanly into a bull camp and a bear camp, and both brought real numbers.

The bull case: the spending is rational and the returns are still there. On The a16z Show (July 14), Gavin Baker, managing partner and chief investment officer of hedge fund Atreides Management, pushed back hard on bubble talk. His central point was that the chips are actually being used: "There are no dark GPUs. All you have to do is read any technical paper and that one of the biggest problems in a trading run is that GPUs melting." (He was contrasting today with the year-2000 telecom bust, when about 97% of the fiber-optic cable laid sat unused, or "dark.") Baker said the biggest spenders "who are all public" have seen "a 10-point increase in their ROICs" since ramping up, ROIC being return on invested capital, a basic measure of whether spending pays off. On valuation, he argued the comparison to the dot-com era is overblown: "Cisco peaked at 150 or 180 times trailing earnings. Nvidia's at more like 40 times. So valuations are very different." His verdict: "I do not believe we're in an AI bubble today."

Baker's a16z colleague David George laid out the buffer behind hyperscaler spending: these companies "collectively generate like $300 billion of free cash flow a year" and hold "$500 billion of cash on the balance sheet," while it costs "like $40 or $50 billion to light up one gigawatt" of data-center capacity. George also cited a startling scale figure: "We have about $1 trillion of data centers in the US. The plan is to add $3 to $4 trillion in the next 5 years... we have already built out in data center capacity a larger amount of dollars than the entire US interstate highway system, which took 40 years." And on usage: "Google released a stat recently that they have seen a 150x increase in the amount of tokens processed in the last 17 months."

On the same show a day later (July 15), Dylan Patel, founder of the respected chip-research firm SemiAnalysis, gave the demand math from the supply side. He said OpenAI and Anthropic together take about 30% of newly deployed chips, that "Meta's CapEx [is] at $60 billion and Google's at $80 billion," and that hyperscalers can "grow CapEx 20-30% next year." He pegged Nvidia's revenue at "over $200 billion" this year, heading to "over $300 billion" next year. But Patel added a genuinely contrarian wrinkle on Nvidia: "the value capture is broken", Nvidia is "not even capturing 10% of the value they've created", and he thinks its pricing power erodes as open-source models and cheaper inference spread.

The bear case: the returns are shrinking and the funding is fragile. The week's marquee bear appearance was Jim Chanos, the famed short-seller who called Enron, on RiskReversal Pod (July 17), in an episode literally titled "The Math Ain't Mathing for the AI Data Center Build." His most important number was about return on incremental invested capital (ROIIC), which measures how much profit each new dollar of spending generates: "Every dollar that the hyperscalers is spending on the margin is on the margin creating less operating income. It's still healthy, but basically it's gone from, as a group, 40% a year and a half ago to about 20% today. And if the spend keeps up at this kind of rate, it's going to be moving toward 10%." His warning on timing: "We're going to get a point in late 26, 27, where... people are going to start putting pencil to paper besides us and say, wait a minute, does this next trillion dollars make sense if you're only going to earn 50 billion on it? We can earn that on treasuries."

Chanos also flagged a duration mismatch, building 10-plus-year assets on the strength of today's prices: "People are making decisions on long-term projects based on spot prices. And that's a terrifying thing. We saw it in the shale business... railroads in the 19th century." And he raised the accounting angle: hyperscalers depreciate chips over "five, six years," but the clock often doesn't start "for 18 months," effectively stretching it to seven-plus years, even though "if you run these chips, 365, 24/7... they burn out after 10 to 12 years."

On RiskReversal Pod a few days earlier (July 10), Louis-Vincent Gave of research firm Gavekal warned the funding model itself is stretched, describing money-losing model companies "burning 10, 15 billion a quarter" as being like a business "selling you the Ferrari at $10,000, it'll be interesting how long the market gets excited about funding people's Ferraris at $10,000." His co-guest Peter Boockvar of One Point BFG noted Oracle "is spending 100% of their revenue on CapEx" and that long-dated data-center leases "are essentially debt."

On the competitive map, one name kept coming up as Nvidia's real rival: Google. Baker was blunt: "Nvidia's biggest competitor, it's not AMD, it's not Broadcom, it's certainly not Marvell, it's not Intel, it's Google. And more specifically... because Google owns the TPU chip" (Google's in-house AI chip). He believes Google may already process more AI traffic than OpenAI or Anthropic, and, contrarian to the popular idea that every big tech firm will build its own chip, he predicted "you'll see a bunch of high-profile ASIC programs canceled" over the next three years if Google starts selling TPUs to outsiders. (An "ASIC" is a custom chip built for one job; Broadcom and Marvell design these for the hyperscalers.)

2. Memory Pricing: HBM, DRAM, NAND (Micron, SK Hynix, Samsung)

Memory was the epicenter this week, both for the money made and the fear that it can't last.

The size of the boom is hard to overstate. On Limitless: An AI Podcast (July 9), the hosts walked through Samsung's quarter: profit of $58.5 billion, which "absolutely crushed Nvidia, who only came in at a measly $53 billion," with roughly 94-96% of Samsung's profit coming from memory alone. They laid out the HBM price increases quarter by quarter: "In Q1, they hiked their prices up 90%... In Q2, it went up a further 50% to 60%... Samsung announced... they're going to hike it up another 20% in Q3." They also explained why HBM squeezes ordinary memory supply: "one gigabyte of HBM consumes the factory capacity of roughly four gigabytes of regular DRAM." And they showed how it reaches consumers: a 32GB RAM kit is "two to three times more expensive this year than it was last year," and Apple raised Mac prices ($200 on the MacBook Air, $1,300 on the Mac Studio) simply "to offset the memory," per Bloomberg's Mark Gurman on Bloomberg Intelligence (July 17).

On Chit Chat Stocks (July 15), host Ryan Henderson gave the jaw-dropping framing: "SK Hynix, Samsung, and Micron in 2026 will generate... more in profits than all of big tech combined... I believe Micron is going to earn more in 2026 than they earned from 2000 to 2025 combined."

The bull argument is that this cycle is different, protected by long contracts. On Power Lunch (July 13), Janus Henderson portfolio manager Kenny Fish made the structural case: memory makers are "for the first time in their history, signing multi-year, long-term supply agreements," which "brackets the range of what the earnings can be through the cycle." She added: "There's going to be no new capacity effectively on for almost two years... supply constrained through 2028," with "a reasonable chance it could be through the end of the decade." On Closing Bell (July 9), Wedbush analyst Matt Bryson said the same shift is real: contracts have moved from the old "3-month" spot deals to "5-year deals" with steep penalties for walking away, and he raised his SanDisk price target to $2,000 from $1,200, targeting gross margins of "84%-85% for SanDisk by 2027."

The bear argument is that memory is still memory, brutally cyclical, and margins this high never last. Peter Boockvar, on Wealthion (July 15), was the clearest: "Never tell me what a PE ratio is on a cyclical stock and tell me it's cheap," because a low price-to-earnings ratio on a cyclical company usually signals a profit peak, not a bargain. On Micron specifically: "Micron has 85% gross margins, and investors are saying, well, how sustainable is an 85% gross margin in what has historically been a highly cyclical space?" His forecast: "You can be sure in the next couple of years this DRAM cycle will reverse. The volume demand will continue up, but the pricing will go down, and you will see a crushing of Micron's earnings." He also noted that only "about 40% [of Micron's revenue] is locked up," leaving 60% exposed to spot prices, and that Micron trades "at about 10 times sales," near where it "peaked at about nine to 10 times sales" in 1999-2000.

On RiskReversal Pod (July 17), Jim Chanos and CNBC's Dan Nathan questioned the durability of Micron's forward order book (its "long-term agreements"): "Why would you believe that... that huge order book, that they are literally being awarded market cap right now for... is going to materialize over the next few years?" Nathan added the pointed question: "If everyone knew that Blackwell and Vera Rubin... was going to need increasingly... expensive memory... why is it that Micron was not bringing on capacity?"

A more nuanced middle view came from Bloomberg Intelligence's Ian King on Bloomberg Tech (July 9): "There's no doubt that demand is very, very strong right now. Nobody is saying that over the next two years or so that we're going to be in anything like balance or oversupply", but he tempered it with the industry's history: "this industry's never got it right over the long term."

3. Semiconductor Capital Equipment / WFE (ASML, AMAT, LRCX, KLAC)

The equipment makers, the companies that sell the machines to build chips, had a genuinely strong week fundamentally, even as their stocks yawned.

ASML was the headliner. On Squawk Box Europe Express (July 15), CNBC reported ASML "raised its 2026 outlook for the second time this year," with Q2 net income of $2.9 billion (versus $2.6B expected) and revenue of $9.3 billion (versus $8.8B), and full-year sales guidance of €43-45 billion. CEO Christophe Fouquet said "customers are continuing to accelerate their expansion plans." On The Canadian Investor (July 16), host Simon Bélanger added the detail that ASML sold "86 new lithography systems, up 28%," that memory and logic sales are now split "half and half" (versus logic outselling memory 2-to-1 in prior years), and that South Korea is now "43-45% of installations", a direct read on the memory build-out. Crucially, ASML plans "a 30% increase in capacity for EUV system in 2027 and possibly another 30% increase in 2028." (EUV, extreme ultraviolet lithography, is ASML's crown-jewel machine; no one else can make one.)

On Bloomberg Tech (July 16), Berenberg's head of tech equity research, Tammy Chu, added that ASML raised its 2026 gross-margin outlook "from midpoint 52% to midpoint 55%," has orders "queued up all the way to 2028," and is building for "every component in this supply chain is under shortage." Her read on the muted stock reaction: "it's just a take a breath," since expectations had "been increasing dramatically."

On the moat and the valuation, Chit Chat Stocks (July 15) went deep. Co-host Brett Schaffer noted ASML's market cap of "$675 billion," a 10-year total return of "34%" a year, EUV machines that sell for "upwards of $350 million apiece," and services revenue growing "18% CAGR since 2012." But he flagged the price: a "PE of 58," near record highs. Notably, both hosts admitted they'd wrongly called ASML "too big to own" a year earlier: Henderson said that in summer 2025, when it "traded down to 25 times earnings... that was probably one of the easiest times to buy. There was no risk to the business model at the time." Schaffer offered a memorable contrast between the equipment makers' restraint and the memory makers' aggression: ASML and TSMC "could charge a billion per machine, but they're not, and it's because they don't want to incentivize people to end their relationship", whereas he thinks the memory names "are being extremely greedy... the stocks are going down 90% at some point."

One useful cross-current from the Chit Chat and Bloomberg discussions: on Bloomberg Tech (July 10), Spear ETF's Ivan Dilevska argued the better risk/reward may be "across semi-cap equipment suppliers (e.g., ASML, Applied Materials, Lam, KLA) rather than just chip makers," because they benefit from everyone's spending regardless of who wins.

4. Foundry / Manufacturing Dynamics (TSM, INTC, GFS)

TSMC's quarter (July 16) was the textbook "beat and fade." On Schwab Network (July 16), FastMarket's Tom White ran the numbers: TSMC "beat on top line revenue... beat on EPS... guided higher," with gross margins that "reached 67.7% this quarter," Q3 revenue guidance of "44.6 billion to 45.8 billion... about 12% sequential growth," and profit margins "up near 56%." And yet the stock fell, because, as Bloomberg's team put it on Bloomberg Tech (July 16), TSMC "said capital expenditures would be higher, not just this year, but for 3 years." The market's new worry, White explained, is "when are we going to see a top in those levels" of capex.

The other big TSMC headline was politics and geography: an additional $100 billion for US factories, bringing total planned US investment to $265 billion. Bloomberg's Mike Sheppard traced the escalation: a "$65 billion investment agreed to under President Joe Biden," raised "to $165 billion soon after Trump took office," then "another $100 billion... under the pressure of tariffs." CEO C.C. Wei has said "it will take years to meet all the demand." On The Financial Exchange Show (July 16), host Mike Armstrong offered a less cynical read than "tariff coercion", a "flywheel" effect: "if you are not building foundries and data centers in the United States, you have something to miss out on." But he was blunt on the reality gap: "There isn't a whole lot of evidence that we are capable of manufacturing the latest and greatest semiconductors here in the United States" yet.

Intel was the week's political-turnaround story. On WSJ Tech News Briefing (July 14), reporter Robbie Whelan explained the Trump administration "converted $9 billion in federal grants... immediately overnight into an equity stake for Intel", the government now owns Intel stock. New CEO Lip-Bu Tan (whom Whelan called "Lipu Tan") has been "poaching really smart engineers... from places like TSMC and also companies like Samsung, SK Hynix." And a wave of customers: "Apple, NVIDIA, and SpaceX have all signed deals with Intel since the government stake was announced." The demand driver is the "agentic boom" needing "hundreds of thousands of... CPUs", Intel's traditional strength. Whelan flagged the risk plainly: "if you put public money into a company like Intel, maybe they don't get it right... that would be a real disaster for everybody because Intel would lose its primary backer."

Dylan Patel (a16z Show, July 15) added an operational diagnosis of Intel's core problem: "it takes them 5 to 6 years to go from design to shipping the product," requiring "14 revisions in some cases" versus an industry standard of "1 to 3." His contrarian prescription was to fix operations rather than immediately split off the factories, and he floated the idea of hyperscalers each investing "$5 billion" to give Intel a "lifeline." He also warned about the world's dependence on one company: TSMC "controls not just all leading edge, [but] the majority of trailing edge production as well... Something needs to be done."

5. Analog / Auto / Industrial Semis (TXN, ADI, MCHP, ON, NXPI, STM)

This was the quietest corner of the podcast world this week. No show offered dedicated analyst or management commentary on Texas Instruments, Analog Devices, Microchip, ON Semiconductor, NXP, or STMicroelectronics. These names came up only in passing, as portfolio holdings or as part of broad "the whole sector sold off" comments. In other words, the week's story was entirely about AI, memory, and foundry, and the more mundane analog and auto-chip world simply wasn't part of the conversation. Worth watching whether that changes as these companies report over the next few weeks (see the watch list below).

6. China / Export Controls / Tariffs

The China angle this week was less about formal export bans and more about competition, and about a couple of specific policy flashpoints.

The market-moving item was an Apple-and-China-memory report. On Bloomberg Intelligence (July 17), analyst Mandeep Singh explained that news of "Apple asking the U.S. government for qualification of a Chinese memory name is creating a lot of downside for the U.S. memory names... your Micron and even SK Hynix." The idea that a marquee American customer might buy Chinese memory rattled the whole group.

Dylan Patel (a16z Show, July 15) gave a detailed, contrarian read on the H20 export debate. (The H20 is the throttled-down Nvidia chip designed to be legal to sell into China.) He argued that selling China H20s is the lesser evil, echoing Nvidia's own case that it helps "stop Huawei from building up a software ecosystem." He also described how Chinese firms already get around restrictions: "Alibaba, Tencent, ByteDance... are renting Blackwell and other chips through a Singaporean company that is effectively a Chinese company." And he sized China's state support of its chip industry at "$150-200 billion a year, through SOEs [state-owned enterprises], through CapEx that's not generating revenue."

On tariffs, the theme was the lingering "echo." On RiskReversal Pod (July 15), FedWatch Advisors' Ben Emons and Dan Nathan tied IBM's warning to tariff after-effects still "reverberating" more than a year later, companies pulling forward orders and stockpiling to beat expected price increases. Peter Boockvar (Wealthion, July 15) made the strategic point that tariffs "force[d] foreign countries and companies to diversify their supply chains" away from US suppliers, which could hurt American firms over time.

7. Earnings Reactions

Earnings season kicked off with a clear pattern: strong numbers, weak stock reactions, and one outright disaster.

The IBM reaction was the one that spooked people most, because of why it fell. On RiskReversal Pod (July 15), Dan Nathan called it "a wake-up call" and a "canary in the coal mine": customers are "double/triple ordering now," which feels great for chip suppliers today but sets up "a very, very dangerous situation in the not so distant future" when that stockpiling reverses. On The Joe Reis Show (July 17), Reis reframed IBM as a story about AI being "substitutional, not additive", "it can destroy revenue pools faster than incumbents can capture new ones."

The recurring lesson across the beats: as Mike Armstrong put it on The Financial Exchange Show (July 16), "You don't get paid as a shareholder for what the company did this quarter. You get paid for what the company's expected to do in the future... expectations were sky high here."

8. Merger / M&A Chatter

There was no blockbuster pure-play chip merger this week, but a few deal threads ran through the podcasts. The most-discussed was the Apple-Broadcom relationship, reported as a roughly $30 billion arrangement, though the coverage (on Market Maker, July 9, and Mac-focused shows) framed it as a supply-chain and geopolitical strategy story rather than a classic acquisition. In chip materials, Chip Stock Investor (July 17) covered Solstice's ~$4.5 billion acquisition of Element Solutions, a consolidation play in the base materials and specialty gases that feed chipmaking; the episode focused narrowly on deal mechanics and the sell-off that followed. And on the memory side, the bigger "deal" energy was around new listings rather than M&A: SK Hynix's Nasdaq debut, the CXMT IPO in Hong Kong (see below), and speculation, relayed by Jim Cramer on Squawk on the Street (July 13), that Samsung will follow SK Hynix with its own US listing: citing Wedbush's Dan Ives, "don't you think Samsung will do it? Oh yeah, Samsung will do it."

9. Cyclicality, Inventory, and the Peak/Trough Debate

This was the intellectual heart of the week, the argument over whether we're near a top.

The "this is a normal shakeout, not a top" camp. On The Financial Exchange Show (July 16), Armstrong Advisory's Chuck Zotta put the moves in perspective with specific price paths: Micron "open[ed] the year at about $300 a share... moved up to $1,200... at $875 right now," and SanDisk went from "$40 a share... [to] $2,500 a share in June... now down 40% to around $1,500." His take: "The semiconductor sector is not going to see its value double every two months forever," and the current wobble is "a little bit of a shaking out." Mandeep Singh (Bloomberg Intelligence, July 17) held that "the fundamentals haven't really changed in a big way" and expects capex guidance to rise at earnings. Ken Fish (Power Lunch, July 13) used a baseball metaphor: the market usually "starts really discounting [a peak]... six to nine months before peak," which "would argue at least for another year, year and a half" of upside.

The "watch out, this looks like a top" camp. On RenMac Off-Script (July 10), RenMac's technical analyst Jeff deGraaf drew the historical parallel: in 2000, semiconductors "peaked two quarters before the first earnings revisions... I'm a little concerned that you've got some of that here." He said RenMac's "bubble indicator... a doubling of the index off of a high in a two year period" had triggered "in April." On Thoughtful Money (July 11), portfolio manager Lance Roberts saw a "pretty defined head and shoulders pattern" forming (a classic topping chart shape) and warned "we're pricing in 2029 earnings right now for a lot of these companies. So the risk is any disappointment." He also raised the demand-destruction worry that these companies "haven't really changed anything about their product except for the price... when supply starts increasing, there's nothing to protect those higher prices."

The valuation-reset framing tied it together. Joe Reis (July 17) captured the consensus of the skeptics without predicting doom: "These semiconductor stocks are priced at perfection right now. They have to deliver... TSMC can report strong results and the sector can still fall because the question is no longer whether demand is strong. The question is whether demand will remain strong enough, concentrated enough, and profitable enough to justify [the] valuations." He also named the deeper risk, an "efficiency paradox": "The hypothesis of more AI usage meant more premium chips, which meant more semiconductor profits, that's no longer guaranteed."

A structural point worth remembering, from Louis-Vincent Gave (RiskReversal Pod, July 10): semiconductors "are now essentially a fifth of the global equity index... it was 2% of the global benchmark" historically. For Asian investors it's even more extreme: "TSMC, SK Hynix and Samsung Electronics... make up 40%" of the Asia benchmark. That concentration is exactly why a chip sell-off can drag whole markets down, and why New York Life's Michael O'Galbo (Market Matters, July 13) noted Korea's market "fell ~18% in past two weeks" after being up more than 70% for the year.

10. China Indigenization (SMIC, CXMT, Huawei) and Impact on US Chip Names

This was thinner than the AI and memory themes, but the commentary that existed was pointed and specific, and it centered on memory.

The recurring warning was that China's memory champions are coming for the Western oligopoly's fat margins. On RiskReversal Pod (July 10), Peter Boockvar was emphatic: "For those that are trading Micron... Have you ever heard of a company called CXMT or Yangtze memory technology?... that 85% gross profit margin that Micron has today, they're coming after it." He returned to it on Wealthion (July 15), noting that "a few days after SK Hynix had their U.S. listing... CXMT in Hong Kong, which is the fourth largest global DRAM producer" did its own IPO, and expressed surprise that "no one's talking about" it. His framing of China's strategy: "China's first focus is market share. It's not profit margin... when Micron's delivering an 85% gross margin... that gross margin is my market share." He added that Yangtze Memory (YMTC) is "more focused on the NAND storage side... coming after SanDisk and Western Digital."

Louis-Vincent Gave (July 10) generalized it into a memorable thesis: "When China enters a space, it gets commoditized... when China enters the room, profits walk out. So it seems to me that we are... very rapidly advancing in the process of commoditizing... the models and the semiconductors. And yet the markets are priced for completely different." A more measured take came from KraneShares' Brandon Ahern on Power Lunch (July 13): CXMT and YMTC are not yet "in the same pool as SK Hynix, Samsung, Micron in terms of high bandwidth memory," but they "are filling a void as the high bandwidth memory makers go upstream... they are gaining a lot of market share" in lower-end applications the incumbents are exiting.

On the AI-model side of indigenization, the Moonshot "Kimi K3" release (see the TL;DR and Section 6) was the concrete catalyst that sent chips lower, a cheaper Chinese model raising the question of whether the West will need to buy quite so many premium chips. As CNBC's Pippa Stevens noted (Squawk on the Street, July 17), Xi Jinping himself used a Beijing AI conference to pitch China as the open, cooperative AI player: AI "should not be a solo performance by any single country, but rather a symphony of global cooperation."

What I'm Watching Next Week

• The heart of chip earnings season. After TSMC and ASML, the market's attention turns to the AI bellwethers and the analog/auto names that were silent this week. Watch for capex guidance from the hyperscalers (Alphabet already guided 2026 capex to $180-190 billion, per The Joe Reis Show, July 17) and whether management teams validate the bulls (numbers keep going up) or the bears (any sign of ROI discipline). Mandeep Singh's bet is that "capex... numbers [will] go up"; the bears are watching for the first crack.
• Whether the IBM "canary" spreads. The key question raised repeatedly this week: is the client stockpiling of memory that hurt IBM a one-off, or the front edge of an inventory glut that eventually hits Micron, SK Hynix, and Samsung? Any commentary about customers pausing orders would be the tell.
• SK Hynix's price stability and a possible Samsung US listing. SK Hynix's chairman said he may do more US share sales "contingent on price stability", and with the stock having fallen back through its $149 debut price, that stability is now in doubt. Watch for confirmation of the widely-predicted Samsung US listing.
• CXMT and Yangtze Memory. Several respected investors flagged these as the most under-appreciated long-term risk to US memory profits. Any share-gain data, HBM progress, or capacity news from China's memory champions is worth tracking.
• The China AI-model cadence. After Moonshot's Kimi K3, the frequency and pricing of new Chinese open-source models is now a direct input into US chip demand expectations. Each cheap, "good enough" release feeds the efficiency-paradox bear case.
• Intel's execution under its government backer. With the US government now an Intel shareholder and Apple/Nvidia/SpaceX signing on, watch for evidence that Intel's manufacturing turnaround is (or isn't) real, the stakes for taxpayers and the sector are unusually high.
• The power bottleneck. A thread running under the chip story: several guests argued the real constraint is now electricity, not silicon. Data-center power deals (natural gas turbines, nuclear) increasingly move chip sentiment.